Contents. ! Data sets. ! Distance and similarity metrics. ! K-means clustering. ! Hierarchical clustering. ! Evaluation of clustering results

Transcription

1 Statistical Analysis of Microarray Data Contents Data sets Distance and similarity metrics K-means clustering Hierarchical clustering Evaluation of clustering results Clustering Jacques van Helden Introduction to clustering Clustering is an unsupervised approach Starting from a set of objects, group them into classes, without any prior description of these classes There are many clustering methods hierarchical k-means self-organizing maps (SOM) knn The results vary drastically depending on clustering method similarity or dissimilarity metric additional parameters specific to each clustering method (e.g. number of centres for the k-mean, agglomeration rule for hierarchical clustering, ) Statistical Analysis of Microarray Data Data sets Jacques van Helden Jacques.van.Helden@ulb.ac.be Diauxic shift DeRisi et al published the first article describing a full-genome monitoring of gene expression data. This article reported an experiment called diauxic shift with with 7 time points. Initially, cells are grown in a glucose-rich medium. As time progresses, cells Consume glucose -> when glucose becomes limiting Glycolysis stops Gluconeogenesis is activated to produce glucose Produce by-products -> the culture medium becomes polluted/ Stress response Cell cycle data Spellman et al. (1998) Time profiles of yeast cells followed during cell cycle. Several experiments were regrouped, with various ways of synchronization (elutriation, cdc mutants, ) ~800 genes showing a periodic patterns of expression were selected (by Fourier analysis) DeRisi et al. Exploring the metabolic and genetic control of gene expression on a genomic scale. Science (1997) vol. 278 (5338) pp Spellman et al. Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Mol Biol Cell (1998) vol. 9 (12) pp

2 Threshold filtering Profiles of regulated genes We applied a chip-wise standardization (centring and scaling) with robust estimates (median and IQR) on each chip. Z-scores obtained after standardization were converted to P-value (normal distribution) to E-value (=P-value*N) Only genes with an E-value < 1 were retained for clustering raffinose.vs.ref.pool ORF YAL066W YAR008W YAR071W YBL005W YBL015W YBL043W YBR018C YBR019C YBR020W raffinose Carbon sources sucrose Z-scores mannose. Chip-wise standardization glucose Gene expression profiles sucrose.vs.ref.pool Filtering Filtering of carbon source data glucose.vs.ref.pool Gasch et al. Genomic expression programs in the response of yeast cells to environmental changes. Mol Biol Cell (2000) vol. 11 (12) pp mannose.vs.ref.pool For the cell cycle experiments, genes had already been filtered in the original publication. We used the 800 selected genes for the analysis. For the diauxic shift and carbon source experiments, each chip contain >6000 genes, most of which are un-regulated. Standardization Gasch et al. (2000), 173 chips (stress response, heat shock, drugs, carbon source, ) We selected the 13 chips with the response to different carbon sources. galactose.vs.ref.pool Data standardization and filtering ethanol.vs.ref.pool Gasch et al. Genomic expression programs in the response of yeast cells to environmental changes. Mol Biol Cell (2000) vol. 11 (12) pp ethanol Gasch et al. (2000), 173 chips (stress response, heat shock, drugs, carbon source, ) galactose Gene expression data - carbon sources fructose.vs.ref.pool Gene expression data response to environmental changes Hierarchical clustering of expression profiles Statistical Analysis of Microarray Data In 1998, Eisen et al. Implemented a software tool called Cluster, which combine hierarchical clustering and heatmap visualization. Applied it to extract clusters of coexpressed genes from various types of expression profiles. Hierarchical clustering Jacques van Helden Jacques.van.Helden@ulb.ac.be Eisen et al. Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci U S A (1998) vol. 95 (25) pp

3 Clustering with gene expression data Hierarchical clustering on gene expression data Gene expression profiles Chip-wise standardization Z-scores Gene expression profiles Chip-wise standardization Z-scores Threshold filtering Profiles of regulated genes Threshold filtering Profiles of regulated genes ethanol galactose glucose mannose. raffinose sucrose Carbon sources ORF YAL066W YAR008W YAR071W YBL005W YBL015W YBL043W YBR018C YBR019C YBR020W ethanol.vs.ref.pool fructose.vs.ref.pool galactose.vs.ref.pool glucose.vs.ref.pool mannose.vs.ref.pool raffinose.vs.ref.pool sucrose.vs.ref.pool Pairwise distance calculation Distance matrix YAL066W YAR008W YAR071W YBL005W YBL015W YBL043W YBR018C YBR019C YBR020W YBR054W YAL066W YAR008W YAR071W YBL005W YBL015W YBL043W YBR018C YBR019C YBR020W YBR054W Hierarchical clustering on gene expression data Gene expression profiles Chip-wise standardization Threshold filtering Pairwise distance calculation Tree building Hierarchical clustering Z-scores Profiles of regulated genes Distance matrix Tree Principle of tree building branch node root Distance matrix Tree representation c4 c3 object 1 c2 c1 object 2 object 3 object 1 object 5 object 4 object 2 object 3 leaf nodes object 4 object 5 object object object object object Hierarchical clustering is an aggregative clustering method takes as input a distance matrix progressively regroups the closest objects/ groups One needs to define a (dis)similarity metric between two groups. There are several possibilities Average linkage: the average distance between objects from groups A and B Single linkage: the distance between the closest objects from groups A and B Complete linkage: the distance between the most distant objects from groups A and B Algorithm (1) Assign each object to a separate cluster. (2) Find the pair of clusters with the shortest distance, and regroup them in a single cluster (3) Repeat (2) until there is a single cluster The result is a tree, whose intermediate nodes represent clusters N objects N-1 intermediate nodes Branch lengths represent distances between clusters Impact of the agglomeration rule The choice of the agglomeration rule has a strong impact on the structure of a tree resulting from hierarchical clustering. Those four trees were built from the same distance matrix, using 4 different agglomeration rules. The clustering order is completely different. Single-linkage typically creates nesting clusters ( Matryoshka dolls ). Complete and Ward linkage create more balanced trees. Note: the matrix was computed from a matrix of random numbers. The subjective impression of structure are thus complete artifacts. Isomorphism on a tree branch node root c4 branch node c4 root c3 c3 c2 c2 c1 c1 leaf 1 leaf 5 leaf 4 leaf 2 leaf 3 leaf 5 leaf 1 leaf 2 leaf 4 leaf 3 In a tree, the two children of any branch node can be swapped. The result is an isomorphic tree, considered as equivalent to the intial one. The two trees shown here are equivalent, however Top tree: leaf 1 is far away from leaf 2 Bottom tree: leaf 1 is neighbour from leaf 2 The vertical distance between two nodes does NOT reflect their actual distance The distance between two nodes is the sum of branch lengths.

4 Golub Impact of the linkage method (Euclidian distance for all the trees) Golub Effect of the distance metrics (complete linkage for all the trees) Golub Gene clustering Golub Ward Biclustering - Euclidian distance Gene clustering highlights groups of genes with similar expression profiles. Single Euclidian Biclustering consists in clustering the rows (genes) and the columns (samples) of the data set. This reveals some subgroups of samples. With the golub 1999 data set The AML and ALL patients are clearly separated at the top level of the tree There are apparently two clusters among the AML samples. Correlation Impact of distance metrics and agglomeration rules Dot product Golub Ward Biclustering - Dot product distance Biclustering consists in clustering the rows (genes) and the columns (samples) of the data set. This reveals some subgroups of samples. With the golub 1999 data set The AML and ALL patients are clearly separated at the top level of the tree There are apparently two clusters among the AML samples. Average Complete Ward

5 Golub Pruning the tree Impact of the linkage method Impact of the distance metric - complete linkage Ipact of the distance metric - single linkage

6 Hierarchical clustering on gene expression data Pruning and cutting the tree The tree can be cut at level k (starting from the root), which creates k clusters Gene expression profiles Chip-wise standardization Z-scores A k-group partitioning is obtained by collecting the leaves below each branch of the pruned tree Threshold filtering Profiles of regulated genes Pairwise distance calculation Distance matrix Hierarchical clustering Tree Tree cut Clusters Statistical Analysis of Microarray Data K-means clustering Clustering around mobile centres The number of centres (k) has to be specified a priori Algorithm (1) Arbitrarily select k initial centres (2) Assign each element to the closest centre (3) Re-calculate centres (mean position of the assigned elements) (4) Repeat (2) and (3) until one of the stopping conditions is reached the clusters are the same as in the previous iteration the difference between two iterations is smaller than a specified threshold the max number of iterations has been reached Jacques van Helden Jacques.van.Helden@ulb.ac.be

7 Mobile centres example - initial conditions Mobile centres example - first iteration Two sets of random points are randomly generated 200 points centred on (0,0) 50 points centred on (1,1) Two points are randomly chosen as seeds (blue dots) Step 1 Each dot is assigned to the cluster with the closest centre Centres are recalculated (blue star) on the basis of the new clusters Mobile centres example - second iteration Mobile centres example - after 3 iterations At each step, points are re-assigned to clusters centres are recalculated Cluster boundaries and centre positions evolve at each iteration At each step, points are re-assigned to clusters centres are recalculated Cluster boundaries and centre positions evolve at each iteration Mobile centres example - after 4 iterations Mobile centres example - after 5 iterations At each step, points are re-assigned to clusters centres are recalculated Cluster boundaries and centre positions evolve at each iteration At each step, points are re-assigned to clusters centres are recalculated Cluster boundaries and centre positions evolve at each iteration

8 Mobile centres example - after 6 iterations Mobile centres example - after 10 iterations At each step, points are re-assigned to clusters centres are recalculated Cluster boundaries and centre positions evolve at each iteration After some iterations (6 in this case), the clusters and centres do not change anymore Mobile centres example - random data K-means clustering K-means clustering is a variant of clustering around mobile centres After each assignation of an element to a centre, the position of this centre is recalculated The convergence is much faster than with the basic mobile centre algorithm after 1 iteration, the result might already be stable K-means is time- and memory-efficient for very large data sets (e.g. thousands of objects) Clustering with gene expression data Clustering can be performed in two ways Taking genes as objects and conditions/cell types as variables Taking conditions/cell types as objects and genes as variables Problem of dimensionality When genes are considered as variables, there are many more variables than objects Generally, only a very small fraction of the genes are regulated (e.g. 30 genes among 6,000) However, all genes will contribute equally to the distance metrics The noise will thus affect the calculated distances between conditions Solution Selection of a subset of strongly regulated genes before applying clustering to conditions/cell types K-means clustering K-means clustering is a variant of clustering around mobile centres After each assignation of an element to a centre, the position of this centre is recalculated The convergence is much faster than with the basic mobile centre algorithm after 1 iteration, the result might already be stable K-means is time- and memory-efficient for very large data sets (e.g. thousands of objects)

9 Diauxic shift: k-means clustering on all genes Diauxic shift: k-means clustering on filtered genes Diauxic shift: k-means clustering on permuted filtered genes Cell cycle data: K-means clustering Cell cycle data: K-means clustering, permuted data Carbon sources: K-means clustering

10 Golub - K-means clustering K-means clustering - summary Strengths Simple to use Fast Can be used with very large data sets Weaknesses The choice of the number of groups is arbitrary The results vary depending on the initial positions of centres The R implementation is based on Euclidian distance, no other metrics are proposed Solutions Try different values for k and compare the result For each value of k, run repeatedly to sample different initial conditions Weakness of the solution Instead of one clustering, you obtain hundreds of different clustering results, totaling thousands of clusters, how to decide among them Statistical Analysis of Microarray Data Evaluation of clustering results How to evaluate the result? It is very hard to make a choice between the multiple possibilities of distance metrics, clustering algorithms and parameters. Several criteria can be used to evaluate the clustering results Consensus: using different methods, comparing the results and extracting a consensus Robustness: running the same algorithm multiple times, with different initial conditions Bootstrap Jack-knife Test different initial positions for the k-means Biological relevance: compare the clustering result to functional annotations (functional catalogs, metabolic pathways, ) Jacques van Helden Jacques.van.Helden@ulb.ac.be Comparing two clustering results Evaluation of robustness - Bootstrap If two methods return partitions of the same size, their clusters can be compared in a confusion table Optimal correspondences between clusters can be established (permuting columns to maximize the diagonal) The consistency between the two classifications can then be estimated with the hit rate Example : Carbon source data, comparison of k-means and hierarchical clustering hierarchical clustering k-means clustering k1 k2 k3 k4 k5 k6 k7 Sum h h h h h h h Sum k-means clustering k4 k3 k5 k1 k2 k7 k6 Sum h h h h h h h Sum Correspondence between clusters hierarchical h1 h2 h3 h4 h5 h6 h7 k-means k4 k3 k5 k1 k2 k7 k6 hierarchical clustering Matches 84 Hit rate Mismatches 49 Error rate 63.2% 36.8% The bootstrap consists in repeating r times (for example r=100) the clustering, using each time Either a different subset of variables Or a different subset of objects The subset of variables is selected randomly, with resampling (i.e. the same variable can be present several times, whilst other variables are absent. On the images the tree is colored according to the reproducibility of the branches during a 100 time bootstrap.